NVIDIA launched the RTX Spark N1X this week, putting its GB10 Arm SoC (system-on-chip) into roughly seven Windows devices at once: Microsoft's Surface Laptop Ultra, another half-dozen laptops, and a mini-PC, the Surface RTX Spark Dev Box.

The move makes NVIDIA the current driving force behind Arm adoption in the Windows ecosystem, a push Microsoft leads with its premium Surface Laptop. NVIDIA has shipped GB10 in Linux systems for around a year, and rumors suggest it wanted to ship even sooner. So the wait for GB10 driving Windows hardware has been long — for NVIDIA, its partners, and end-users alike.

What does this change for buyers? Choice, mainly. Some manufacturers are following Microsoft's line and targeting developers and content creators, consistent with what GB10 has shown so far. Others have built laptops for the broader consumer market. NVIDIA positions the RTX Spark as a high-end SoC in the vein of AMD's Ryzen AI Max series (Strix Halo), but pitched at a wider user base than Microsoft's laptop alone.

What's inside the N1X?

The N1X is the consumer-facing name for GB10, and the branding leaves room for future RTX Spark family chips. NVIDIA detailed the architecture at Hot Chips 2025, and most of it still applies.

GB10 is a joint project between NVIDIA and MediaTek, built from two chiplets:

• S-chiplet (MediaTek): CPU cores, memory controllers, display controllers, and general-purpose I/O including the PCIe lanes
• G-chiplet (NVIDIA): the graphics engine

Both chiplets are fabbed on TSMC's 3nm process, which at GB10's launch was the fab's leading node — newer than the node behind NVIDIA's Blackwell GPUs or its Grace CPU. By that measure, GB10 is the most advanced Blackwell-architecture chip NVIDIA has designed.

How does the CPU side hold up?

The MediaTek die packs 10 Arm Cortex-X925 performance cores and 10 Cortex-A725 efficiency cores, organized in two clusters of two core groups each. Each cluster carries its own 16MB of L3 cache. These are off-the-shelf cores from Arm's 2024 Cortex lineup, not derived from NVIDIA's Grace.

That timing matters. Launching a chip built on 2024 CPU cores in 2026 puts the N1X behind the curve in CPU technology, with real performance repercussions. The design emphasizes multi-threaded throughput over single-threaded speed and doesn't push high clock speeds.

The memory subsystem is the strongest part of the story. MediaTek fitted a 16-channel, 256-bit LPDDR5X bus — one of just a handful of non-Apple consumer chips with a bus that wide. It delivers 300GB/second of bandwidth on LPDDR5X-9400 and supports up to 128GB of memory in the current configuration.

What about the GPU and the chip-to-chip link?

The G-chiplet is a 48-SM (streaming multiprocessor) Blackwell implementation, 6144 CUDA cores in total. In hardware terms that's equivalent to a GeForce RTX 5070, though the discrete card gets more than twice the memory bandwidth from GDDR7.

It's a full Blackwell graphics design, not a stripped-down server derivative. That means ray tracing hardware, texture units, ROPs (render output units) — everything needed to draw a pixel, plus tensor cores. NVIDIA expects support for its full graphics stack, including DLSS (its AI upscaling technology), to be a big differentiator in the consumer space.

The two chiplets connect over NVLink-C2C, NVIDIA's chip-to-chip interface, at 600GB/second — roughly twice the speed of the memory controllers. The same technology links Grace CPUs to Blackwell GPUs in server-grade GB200/GB300 systems.

One floorplan quirk: the NVIDIA-designed display controllers sit on the MediaTek die, next to the GPU hardware's absence notwithstanding, because display controllers perform best close to memory controllers for bandwidth and power reasons. They sit directly beside one memory controller block.

NVIDIA has called GB10/N1X a successful collaboration with MediaTek. Whether the partnership repeats is up in the air. But after 15 years of solo Arm SoC design across Tegra, DRIVE, and Jetson, a non-monolithic, co-designed chip from NVIDIA is notable in itself.